brintos

brintos / linux-shallow public Read only

0
0
Text · 17.1 KiB · 8780aa2 Raw
677 lines · c
1/*2 * SPDX-License-Identifier: MIT3 *4 * Copyright © 2014-2016 Intel Corporation5 */6 7#include <drm/drm_cache.h>8#include <linux/vmalloc.h>9 10#include "gt/intel_gt.h"11#include "gt/intel_tlb.h"12 13#include "i915_drv.h"14#include "i915_gem_object.h"15#include "i915_scatterlist.h"16#include "i915_gem_lmem.h"17#include "i915_gem_mman.h"18 19void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj,20				 struct sg_table *pages)21{22	struct drm_i915_private *i915 = to_i915(obj->base.dev);23	unsigned long supported = RUNTIME_INFO(i915)->page_sizes;24	bool shrinkable;25	int i;26 27	assert_object_held_shared(obj);28 29	if (i915_gem_object_is_volatile(obj))30		obj->mm.madv = I915_MADV_DONTNEED;31 32	/* Make the pages coherent with the GPU (flushing any swapin). */33	if (obj->cache_dirty) {34		WARN_ON_ONCE(IS_DGFX(i915));35		obj->write_domain = 0;36		if (i915_gem_object_has_struct_page(obj))37			drm_clflush_sg(pages);38		obj->cache_dirty = false;39	}40 41	obj->mm.get_page.sg_pos = pages->sgl;42	obj->mm.get_page.sg_idx = 0;43	obj->mm.get_dma_page.sg_pos = pages->sgl;44	obj->mm.get_dma_page.sg_idx = 0;45 46	obj->mm.pages = pages;47 48	obj->mm.page_sizes.phys = i915_sg_dma_sizes(pages->sgl);49	GEM_BUG_ON(!obj->mm.page_sizes.phys);50 51	/*52	 * Calculate the supported page-sizes which fit into the given53	 * sg_page_sizes. This will give us the page-sizes which we may be able54	 * to use opportunistically when later inserting into the GTT. For55	 * example if phys=2G, then in theory we should be able to use 1G, 2M,56	 * 64K or 4K pages, although in practice this will depend on a number of57	 * other factors.58	 */59	obj->mm.page_sizes.sg = 0;60	for_each_set_bit(i, &supported, ilog2(I915_GTT_MAX_PAGE_SIZE) + 1) {61		if (obj->mm.page_sizes.phys & ~0u << i)62			obj->mm.page_sizes.sg |= BIT(i);63	}64	GEM_BUG_ON(!HAS_PAGE_SIZES(i915, obj->mm.page_sizes.sg));65 66	shrinkable = i915_gem_object_is_shrinkable(obj);67 68	if (i915_gem_object_is_tiled(obj) &&69	    i915->gem_quirks & GEM_QUIRK_PIN_SWIZZLED_PAGES) {70		GEM_BUG_ON(i915_gem_object_has_tiling_quirk(obj));71		i915_gem_object_set_tiling_quirk(obj);72		GEM_BUG_ON(!list_empty(&obj->mm.link));73		atomic_inc(&obj->mm.shrink_pin);74		shrinkable = false;75	}76 77	if (shrinkable && !i915_gem_object_has_self_managed_shrink_list(obj)) {78		struct list_head *list;79		unsigned long flags;80 81		assert_object_held(obj);82		spin_lock_irqsave(&i915->mm.obj_lock, flags);83 84		i915->mm.shrink_count++;85		i915->mm.shrink_memory += obj->base.size;86 87		if (obj->mm.madv != I915_MADV_WILLNEED)88			list = &i915->mm.purge_list;89		else90			list = &i915->mm.shrink_list;91		list_add_tail(&obj->mm.link, list);92 93		atomic_set(&obj->mm.shrink_pin, 0);94		spin_unlock_irqrestore(&i915->mm.obj_lock, flags);95	}96}97 98int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj)99{100	struct drm_i915_private *i915 = to_i915(obj->base.dev);101	int err;102 103	assert_object_held_shared(obj);104 105	if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) {106		drm_dbg(&i915->drm,107			"Attempting to obtain a purgeable object\n");108		return -EFAULT;109	}110 111	err = obj->ops->get_pages(obj);112	GEM_BUG_ON(!err && !i915_gem_object_has_pages(obj));113 114	return err;115}116 117/* Ensure that the associated pages are gathered from the backing storage118 * and pinned into our object. i915_gem_object_pin_pages() may be called119 * multiple times before they are released by a single call to120 * i915_gem_object_unpin_pages() - once the pages are no longer referenced121 * either as a result of memory pressure (reaping pages under the shrinker)122 * or as the object is itself released.123 */124int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj)125{126	int err;127 128	assert_object_held(obj);129 130	assert_object_held_shared(obj);131 132	if (unlikely(!i915_gem_object_has_pages(obj))) {133		GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));134 135		err = ____i915_gem_object_get_pages(obj);136		if (err)137			return err;138 139		smp_mb__before_atomic();140	}141	atomic_inc(&obj->mm.pages_pin_count);142 143	return 0;144}145 146int i915_gem_object_pin_pages_unlocked(struct drm_i915_gem_object *obj)147{148	struct i915_gem_ww_ctx ww;149	int err;150 151	i915_gem_ww_ctx_init(&ww, true);152retry:153	err = i915_gem_object_lock(obj, &ww);154	if (!err)155		err = i915_gem_object_pin_pages(obj);156 157	if (err == -EDEADLK) {158		err = i915_gem_ww_ctx_backoff(&ww);159		if (!err)160			goto retry;161	}162	i915_gem_ww_ctx_fini(&ww);163	return err;164}165 166/* Immediately discard the backing storage */167int i915_gem_object_truncate(struct drm_i915_gem_object *obj)168{169	if (obj->ops->truncate)170		return obj->ops->truncate(obj);171 172	return 0;173}174 175static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj)176{177	struct radix_tree_iter iter;178	void __rcu **slot;179 180	rcu_read_lock();181	radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0)182		radix_tree_delete(&obj->mm.get_page.radix, iter.index);183	radix_tree_for_each_slot(slot, &obj->mm.get_dma_page.radix, &iter, 0)184		radix_tree_delete(&obj->mm.get_dma_page.radix, iter.index);185	rcu_read_unlock();186}187 188static void unmap_object(struct drm_i915_gem_object *obj, void *ptr)189{190	if (is_vmalloc_addr(ptr))191		vunmap(ptr);192}193 194static void flush_tlb_invalidate(struct drm_i915_gem_object *obj)195{196	struct drm_i915_private *i915 = to_i915(obj->base.dev);197	struct intel_gt *gt;198	int id;199 200	for_each_gt(gt, i915, id) {201		if (!obj->mm.tlb[id])202			continue;203 204		intel_gt_invalidate_tlb_full(gt, obj->mm.tlb[id]);205		obj->mm.tlb[id] = 0;206	}207}208 209struct sg_table *210__i915_gem_object_unset_pages(struct drm_i915_gem_object *obj)211{212	struct sg_table *pages;213 214	assert_object_held_shared(obj);215 216	pages = fetch_and_zero(&obj->mm.pages);217	if (IS_ERR_OR_NULL(pages))218		return pages;219 220	if (i915_gem_object_is_volatile(obj))221		obj->mm.madv = I915_MADV_WILLNEED;222 223	if (!i915_gem_object_has_self_managed_shrink_list(obj))224		i915_gem_object_make_unshrinkable(obj);225 226	if (obj->mm.mapping) {227		unmap_object(obj, page_mask_bits(obj->mm.mapping));228		obj->mm.mapping = NULL;229	}230 231	__i915_gem_object_reset_page_iter(obj);232	obj->mm.page_sizes.phys = obj->mm.page_sizes.sg = 0;233 234	flush_tlb_invalidate(obj);235 236	return pages;237}238 239int __i915_gem_object_put_pages(struct drm_i915_gem_object *obj)240{241	struct sg_table *pages;242 243	if (i915_gem_object_has_pinned_pages(obj))244		return -EBUSY;245 246	/* May be called by shrinker from within get_pages() (on another bo) */247	assert_object_held_shared(obj);248 249	i915_gem_object_release_mmap_offset(obj);250 251	/*252	 * ->put_pages might need to allocate memory for the bit17 swizzle253	 * array, hence protect them from being reaped by removing them from gtt254	 * lists early.255	 */256	pages = __i915_gem_object_unset_pages(obj);257 258	/*259	 * XXX Temporary hijinx to avoid updating all backends to handle260	 * NULL pages. In the future, when we have more asynchronous261	 * get_pages backends we should be better able to handle the262	 * cancellation of the async task in a more uniform manner.263	 */264	if (!IS_ERR_OR_NULL(pages))265		obj->ops->put_pages(obj, pages);266 267	return 0;268}269 270/* The 'mapping' part of i915_gem_object_pin_map() below */271static void *i915_gem_object_map_page(struct drm_i915_gem_object *obj,272				      enum i915_map_type type)273{274	unsigned long n_pages = obj->base.size >> PAGE_SHIFT, i;275	struct page *stack[32], **pages = stack, *page;276	struct sgt_iter iter;277	pgprot_t pgprot;278	void *vaddr;279 280	switch (type) {281	default:282		MISSING_CASE(type);283		fallthrough;	/* to use PAGE_KERNEL anyway */284	case I915_MAP_WB:285		/*286		 * On 32b, highmem using a finite set of indirect PTE (i.e.287		 * vmap) to provide virtual mappings of the high pages.288		 * As these are finite, map_new_virtual() must wait for some289		 * other kmap() to finish when it runs out. If we map a large290		 * number of objects, there is no method for it to tell us291		 * to release the mappings, and we deadlock.292		 *293		 * However, if we make an explicit vmap of the page, that294		 * uses a larger vmalloc arena, and also has the ability295		 * to tell us to release unwanted mappings. Most importantly,296		 * it will fail and propagate an error instead of waiting297		 * forever.298		 *299		 * So if the page is beyond the 32b boundary, make an explicit300		 * vmap.301		 */302		if (n_pages == 1 && !PageHighMem(sg_page(obj->mm.pages->sgl)))303			return page_address(sg_page(obj->mm.pages->sgl));304		pgprot = PAGE_KERNEL;305		break;306	case I915_MAP_WC:307		pgprot = pgprot_writecombine(PAGE_KERNEL_IO);308		break;309	}310 311	if (n_pages > ARRAY_SIZE(stack)) {312		/* Too big for stack -- allocate temporary array instead */313		pages = kvmalloc_array(n_pages, sizeof(*pages), GFP_KERNEL);314		if (!pages)315			return ERR_PTR(-ENOMEM);316	}317 318	i = 0;319	for_each_sgt_page(page, iter, obj->mm.pages)320		pages[i++] = page;321	vaddr = vmap(pages, n_pages, 0, pgprot);322	if (pages != stack)323		kvfree(pages);324 325	return vaddr ?: ERR_PTR(-ENOMEM);326}327 328static void *i915_gem_object_map_pfn(struct drm_i915_gem_object *obj,329				     enum i915_map_type type)330{331	resource_size_t iomap = obj->mm.region->iomap.base -332		obj->mm.region->region.start;333	unsigned long n_pfn = obj->base.size >> PAGE_SHIFT;334	unsigned long stack[32], *pfns = stack, i;335	struct sgt_iter iter;336	dma_addr_t addr;337	void *vaddr;338 339	GEM_BUG_ON(type != I915_MAP_WC);340 341	if (n_pfn > ARRAY_SIZE(stack)) {342		/* Too big for stack -- allocate temporary array instead */343		pfns = kvmalloc_array(n_pfn, sizeof(*pfns), GFP_KERNEL);344		if (!pfns)345			return ERR_PTR(-ENOMEM);346	}347 348	i = 0;349	for_each_sgt_daddr(addr, iter, obj->mm.pages)350		pfns[i++] = (iomap + addr) >> PAGE_SHIFT;351	vaddr = vmap_pfn(pfns, n_pfn, pgprot_writecombine(PAGE_KERNEL_IO));352	if (pfns != stack)353		kvfree(pfns);354 355	return vaddr ?: ERR_PTR(-ENOMEM);356}357 358/* get, pin, and map the pages of the object into kernel space */359void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,360			      enum i915_map_type type)361{362	enum i915_map_type has_type;363	bool pinned;364	void *ptr;365	int err;366 367	if (!i915_gem_object_has_struct_page(obj) &&368	    !i915_gem_object_has_iomem(obj))369		return ERR_PTR(-ENXIO);370 371	if (WARN_ON_ONCE(obj->flags & I915_BO_ALLOC_GPU_ONLY))372		return ERR_PTR(-EINVAL);373 374	assert_object_held(obj);375 376	pinned = !(type & I915_MAP_OVERRIDE);377	type &= ~I915_MAP_OVERRIDE;378 379	if (!atomic_inc_not_zero(&obj->mm.pages_pin_count)) {380		if (unlikely(!i915_gem_object_has_pages(obj))) {381			GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj));382 383			err = ____i915_gem_object_get_pages(obj);384			if (err)385				return ERR_PTR(err);386 387			smp_mb__before_atomic();388		}389		atomic_inc(&obj->mm.pages_pin_count);390		pinned = false;391	}392	GEM_BUG_ON(!i915_gem_object_has_pages(obj));393 394	/*395	 * For discrete our CPU mappings needs to be consistent in order to396	 * function correctly on !x86. When mapping things through TTM, we use397	 * the same rules to determine the caching type.398	 *399	 * The caching rules, starting from DG1:400	 *401	 *	- If the object can be placed in device local-memory, then the402	 *	  pages should be allocated and mapped as write-combined only.403	 *404	 *	- Everything else is always allocated and mapped as write-back,405	 *	  with the guarantee that everything is also coherent with the406	 *	  GPU.407	 *408	 * Internal users of lmem are already expected to get this right, so no409	 * fudging needed there.410	 */411	if (i915_gem_object_placement_possible(obj, INTEL_MEMORY_LOCAL)) {412		if (type != I915_MAP_WC && !obj->mm.n_placements) {413			ptr = ERR_PTR(-ENODEV);414			goto err_unpin;415		}416 417		type = I915_MAP_WC;418	} else if (IS_DGFX(to_i915(obj->base.dev))) {419		type = I915_MAP_WB;420	}421 422	ptr = page_unpack_bits(obj->mm.mapping, &has_type);423	if (ptr && has_type != type) {424		if (pinned) {425			ptr = ERR_PTR(-EBUSY);426			goto err_unpin;427		}428 429		unmap_object(obj, ptr);430 431		ptr = obj->mm.mapping = NULL;432	}433 434	if (!ptr) {435		err = i915_gem_object_wait_moving_fence(obj, true);436		if (err) {437			ptr = ERR_PTR(err);438			goto err_unpin;439		}440 441		if (GEM_WARN_ON(type == I915_MAP_WC && !pat_enabled()))442			ptr = ERR_PTR(-ENODEV);443		else if (i915_gem_object_has_struct_page(obj))444			ptr = i915_gem_object_map_page(obj, type);445		else446			ptr = i915_gem_object_map_pfn(obj, type);447		if (IS_ERR(ptr))448			goto err_unpin;449 450		obj->mm.mapping = page_pack_bits(ptr, type);451	}452 453	return ptr;454 455err_unpin:456	atomic_dec(&obj->mm.pages_pin_count);457	return ptr;458}459 460void *i915_gem_object_pin_map_unlocked(struct drm_i915_gem_object *obj,461				       enum i915_map_type type)462{463	void *ret;464 465	i915_gem_object_lock(obj, NULL);466	ret = i915_gem_object_pin_map(obj, type);467	i915_gem_object_unlock(obj);468 469	return ret;470}471 472void __i915_gem_object_flush_map(struct drm_i915_gem_object *obj,473				 unsigned long offset,474				 unsigned long size)475{476	enum i915_map_type has_type;477	void *ptr;478 479	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));480	GEM_BUG_ON(range_overflows_t(typeof(obj->base.size),481				     offset, size, obj->base.size));482 483	wmb(); /* let all previous writes be visible to coherent partners */484	obj->mm.dirty = true;485 486	if (obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE)487		return;488 489	ptr = page_unpack_bits(obj->mm.mapping, &has_type);490	if (has_type == I915_MAP_WC)491		return;492 493	drm_clflush_virt_range(ptr + offset, size);494	if (size == obj->base.size) {495		obj->write_domain &= ~I915_GEM_DOMAIN_CPU;496		obj->cache_dirty = false;497	}498}499 500void __i915_gem_object_release_map(struct drm_i915_gem_object *obj)501{502	GEM_BUG_ON(!obj->mm.mapping);503 504	/*505	 * We allow removing the mapping from underneath pinned pages!506	 *507	 * Furthermore, since this is an unsafe operation reserved only508	 * for construction time manipulation, we ignore locking prudence.509	 */510	unmap_object(obj, page_mask_bits(fetch_and_zero(&obj->mm.mapping)));511 512	i915_gem_object_unpin_map(obj);513}514 515struct scatterlist *516__i915_gem_object_page_iter_get_sg(struct drm_i915_gem_object *obj,517				   struct i915_gem_object_page_iter *iter,518				   pgoff_t n,519				   unsigned int *offset)520 521{522	const bool dma = iter == &obj->mm.get_dma_page ||523			 iter == &obj->ttm.get_io_page;524	unsigned int idx, count;525	struct scatterlist *sg;526 527	might_sleep();528	GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT);529	if (!i915_gem_object_has_pinned_pages(obj))530		assert_object_held(obj);531 532	/* As we iterate forward through the sg, we record each entry in a533	 * radixtree for quick repeated (backwards) lookups. If we have seen534	 * this index previously, we will have an entry for it.535	 *536	 * Initial lookup is O(N), but this is amortized to O(1) for537	 * sequential page access (where each new request is consecutive538	 * to the previous one). Repeated lookups are O(lg(obj->base.size)),539	 * i.e. O(1) with a large constant!540	 */541	if (n < READ_ONCE(iter->sg_idx))542		goto lookup;543 544	mutex_lock(&iter->lock);545 546	/* We prefer to reuse the last sg so that repeated lookup of this547	 * (or the subsequent) sg are fast - comparing against the last548	 * sg is faster than going through the radixtree.549	 */550 551	sg = iter->sg_pos;552	idx = iter->sg_idx;553	count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg);554 555	while (idx + count <= n) {556		void *entry;557		unsigned long i;558		int ret;559 560		/* If we cannot allocate and insert this entry, or the561		 * individual pages from this range, cancel updating the562		 * sg_idx so that on this lookup we are forced to linearly563		 * scan onwards, but on future lookups we will try the564		 * insertion again (in which case we need to be careful of565		 * the error return reporting that we have already inserted566		 * this index).567		 */568		ret = radix_tree_insert(&iter->radix, idx, sg);569		if (ret && ret != -EEXIST)570			goto scan;571 572		entry = xa_mk_value(idx);573		for (i = 1; i < count; i++) {574			ret = radix_tree_insert(&iter->radix, idx + i, entry);575			if (ret && ret != -EEXIST)576				goto scan;577		}578 579		idx += count;580		sg = ____sg_next(sg);581		count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg);582	}583 584scan:585	iter->sg_pos = sg;586	iter->sg_idx = idx;587 588	mutex_unlock(&iter->lock);589 590	if (unlikely(n < idx)) /* insertion completed by another thread */591		goto lookup;592 593	/* In case we failed to insert the entry into the radixtree, we need594	 * to look beyond the current sg.595	 */596	while (idx + count <= n) {597		idx += count;598		sg = ____sg_next(sg);599		count = dma ? __sg_dma_page_count(sg) : __sg_page_count(sg);600	}601 602	*offset = n - idx;603	return sg;604 605lookup:606	rcu_read_lock();607 608	sg = radix_tree_lookup(&iter->radix, n);609	GEM_BUG_ON(!sg);610 611	/* If this index is in the middle of multi-page sg entry,612	 * the radix tree will contain a value entry that points613	 * to the start of that range. We will return the pointer to614	 * the base page and the offset of this page within the615	 * sg entry's range.616	 */617	*offset = 0;618	if (unlikely(xa_is_value(sg))) {619		unsigned long base = xa_to_value(sg);620 621		sg = radix_tree_lookup(&iter->radix, base);622		GEM_BUG_ON(!sg);623 624		*offset = n - base;625	}626 627	rcu_read_unlock();628 629	return sg;630}631 632struct page *633__i915_gem_object_get_page(struct drm_i915_gem_object *obj, pgoff_t n)634{635	struct scatterlist *sg;636	unsigned int offset;637 638	GEM_BUG_ON(!i915_gem_object_has_struct_page(obj));639 640	sg = i915_gem_object_get_sg(obj, n, &offset);641	return nth_page(sg_page(sg), offset);642}643 644/* Like i915_gem_object_get_page(), but mark the returned page dirty */645struct page *646__i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, pgoff_t n)647{648	struct page *page;649 650	page = i915_gem_object_get_page(obj, n);651	if (!obj->mm.dirty)652		set_page_dirty(page);653 654	return page;655}656 657dma_addr_t658__i915_gem_object_get_dma_address_len(struct drm_i915_gem_object *obj,659				      pgoff_t n, unsigned int *len)660{661	struct scatterlist *sg;662	unsigned int offset;663 664	sg = i915_gem_object_get_sg_dma(obj, n, &offset);665 666	if (len)667		*len = sg_dma_len(sg) - (offset << PAGE_SHIFT);668 669	return sg_dma_address(sg) + (offset << PAGE_SHIFT);670}671 672dma_addr_t673__i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj, pgoff_t n)674{675	return i915_gem_object_get_dma_address_len(obj, n, NULL);676}677